Feature: Training a pre-trained (pre-created) model

[rjzak]

I know that training a model is not part of the goals of the project, but I'd like to suggest the ability to train a pre-trained model.

This would allow training, but not creating the model. It seems that the reason for not allowing training is the difficulty in creating the model structure itself with all of the various parameters and options, which may be too much for a Wasi application. However, evaluating a model isn't much different from training, which just updates the weights based on error. Maybe it could be a nice medium between no training at all vs. just updating weights but no model creation?

[abrown]

cc: @mingqiusun

[mingqiusun]

@rjzak Adding training is always on the back of our minds. In order to support training, we need to expose cost function, variable set, and back propagation. So it is more complicated than just inferencing, and different backends may do things differently. But if you want to propose an API for doing that, let's have a discussion.

[tonibofarull]

Will it be possible to use the subset of functions for inference without having to implement the training ones?